JP6876589B2 - Anomaly detection device, anomaly detection method, and anomaly detection program - Google Patents

Description

The present invention relates to an abnormality detection device, an abnormality detection method, and an abnormality detection program that analyze waveforms observed by various sensors such as a microphone, a vibration sensor, and an optical sensor and detect an abnormality from the change.

For example, when detecting equipment abnormalities such as machine vibration and operating noise, belt conveyor vibration, and bearing wear, waveforms observed by various sensors such as microphones, vibration sensors, and optical sensors are analyzed, and abnormalities are detected from the changes. In order to detect, it is necessary to analyze a wide range of signals from electrical or mechanical high-speed transient phenomena to low-speed repetitive phenomena.

Conventionally, when detecting an abnormality in a signal to be measured from various sensors of this type, a plurality of types of feature quantities in which the characteristics of the signal to be measured are quantified and a threshold value for each feature quantity are designed in advance and measured. Multiple types of feature quantities are extracted from the digital data of the signal, each extracted feature quantity is compared with the threshold value, and if any of the feature quantities exceeds the threshold value from these comparison results, the signal to be measured is regarded as abnormal. It was common to judge. For example, only a predetermined signal component of the signal to be measured converted into digital data is extracted by a digital filter, the peak value of the extracted signal component is extracted as a feature amount, and the extracted feature amount (peak value) is equal to or higher than the threshold value. If the signal to be measured is judged to be abnormal.

Here, in Patent Document 1 below, as a method using a feature amount, the peak wavelength of the spectral distribution of the reflected light in the measurement region of the measurement object is extracted as the feature amount, and the extracted feature amount and the feature acquired in advance are described. A displacement measuring method for calculating the displacement of a measurement region based on the relationship between the quantity and the displacement is disclosed.

Japanese Patent No. 5701837

By the way, the characteristics of regular signals such as sine waves and square waves can be easily analyzed with an oscilloscope, a digital meter, and a frequency meter. However, there is a problem that analysis of irregular signals including sudden mechanical vibrations, lightning surges, noises, single signals such as sensor signals, noises, distortions, etc. requires extremely complicated calculation work.

Then, in the method of extracting the above-mentioned plurality of types of feature amounts and comparing each extracted feature amount with the corresponding threshold value, there is a risk of causing an erroneous determination because the threshold value itself varies. Further, the method of extracting the feature amount and comparing it with the threshold value has a problem that it is specialized in the signal to be analyzed and has low versatility.

In addition, when applied to product inspections on production lines, etc., if an attempt is made to improve the abnormality detection accuracy by loosely setting the threshold value for each feature amount, there is a risk that even a non-defective product that should be judged to be normal may be erroneously judged as abnormal. there were. Moreover, if the number of erroneous judgments increases, the rate of wasting non-defective products also increases, and the yield of products may deteriorate, resulting in an extremely low production efficiency.

For this reason, it is necessary to constantly monitor the state of monitoring equipment and manufacturing equipment, and it has been desired to provide an abnormality detection device capable of detecting an abnormality more accurately with as few false alarms as possible.

Therefore, the present invention has been made in view of the above problems, and an abnormality detection device, an abnormality detection method, and an abnormality detection program capable of detecting an abnormality more accurately than before without depending on an analysis target. It is intended to be provided.

In order to achieve the above object, the abnormality detection device according to claim 1 of the present invention includes a data storage unit 4 for storing digital data of a signal to be measured and a data storage unit 4.
An abnormality detection device 1 including an identification unit 5 for identifying whether the digital data is normal or abnormal.
The identification unit
A digital signal processor 11 that extracts a predetermined feature amount from the digital data stored in the data storage unit, and a digital signal processor 11.
A provisional determination device 12 that provisionally determines whether or not the digital data is normal based on the extraction result of the feature amount by the digital signal processor, and
A machine learning classifier 13 that discriminates whether digital data that is tentatively determined to be normal by the tentative determination device is normal or abnormal by machine learning ,
A teacher data storage unit 17 that stores teacher data labeled as normal or abnormal, and
A classifier tuning data storage unit 15 that stores classifier tuning data labeled as normal or abnormal, and a classifier tuning data storage unit 15.
The output value of the machine learning classifier of the classifier tuning data with the normal label of the classifier tuning data storage unit and the machine learning classifier of the digital data tentatively determined to be normal by the tentative judgment device. The classifier tuning data update unit 14 that updates the normally labeled classifier tuning data stored in the classifier tuning data storage unit based on the comparison result with the output value, and the classifier tuning data update unit 14.
Among the updated normal labeled classifier tuning data of the classifier tuning data storage unit, the classifier tuning data in which the output value of the machine learning classifier is equal to or higher than a preset threshold value is normal. The teacher data update unit 16 which is additionally stored in the teacher data storage unit as the teacher data labeled with
It is characterized by including a machine learning classifier updating unit 18 that updates the machine learning classifier using the teacher data with a normal label updated by the teacher data updating unit.

The abnormality detection method according to claim 2 is an abnormality detection method that stores digital data of a signal to be measured and identifies whether the digital data is normal or abnormal.
A step of extracting a predetermined feature amount from the stored digital data, and
A step of tentatively determining whether or not the digital data is normal based on the extraction result of the feature amount, and
A step of identifying whether the digital data tentatively determined to be normal by the tentative determination process is normal or abnormal by machine learning , and
Steps to store teacher data labeled normal or abnormal, and
Steps to store classifier tuning data labeled normal or abnormal, and
Normal label based on the comparison result between the output value of the classifier tuning data with a normal label by the machine learning and the output value of the digital data tentatively determined to be normal by the tentative determination process by the machine learning. Steps to update the classifier tuning data with
Of the updated classifier tuning data with a normal label, the classifier tuning data whose output value by machine learning is equal to or higher than a preset threshold value is additionally stored as teacher data with a normal label. Steps and
It is characterized by including a step of updating the machine learning with the updated normal labeled teacher data.

The abnormality detection program according to claim 3 uses a computer.
A data storage unit 4 that stores digital data of the signal to be measured, and
A digital signal processor 11 that extracts a predetermined feature amount from the digital data stored in the data storage unit, and a digital signal processor 11.
A provisional determination device 12 that provisionally determines whether or not the digital data is normal based on the extraction result of the feature amount by the digital signal processor, and
A machine learning classifier 13 that discriminates whether digital data that is tentatively determined to be normal by the tentative determination device is normal or abnormal by machine learning ,
A teacher data storage unit 17 that stores teacher data labeled as normal or abnormal, and
A classifier tuning data storage unit 15 that stores classifier tuning data labeled as normal or abnormal, and a classifier tuning data storage unit 15.
The output value of the machine learning classifier of the classifier tuning data with the normal label of the classifier tuning data storage unit and the machine learning classifier of the digital data tentatively determined to be normal by the tentative judgment device. The classifier tuning data update unit 14 that updates the normally labeled classifier tuning data stored in the classifier tuning data storage unit based on the comparison result with the output value, and the classifier tuning data update unit 14.
Among the updated normal labeled classifier tuning data of the classifier tuning data storage unit, the classifier tuning data in which the output value of the machine learning classifier is equal to or higher than a preset threshold value is normal. The teacher data update unit 16 which is additionally stored in the teacher data storage unit as the teacher data labeled with
It is characterized in that it functions as a machine learning classifier update unit 18 that updates the machine learning classifier using the teacher data with a normal label updated by the teacher data update unit.

According to the present invention, even if the signal to be measured is a regular signal, a sudden signal, or an irregular signal, it is possible to detect an abnormality more accurately than before without depending on the analysis target.

It is a block block diagram of the abnormality detection apparatus which concerns on this invention. It is a flowchart at the time of setting the identification boundary value of a machine learning classifier. It is a flowchart at the time of updating a machine learning classifier. It is a flowchart at the time of updating the identification boundary value of a machine learning classifier. It is a flowchart at the time of operation.

Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the attached drawings.

First, definitions of terms used in this embodiment will be described. The "abnormal" is a state different from the normal state, and means a state in which the value deviates from the normal state. In addition, "anomaly detection" means finding data that behaves differently from the majority of other data, finding unique data compared to others, finding unusual patterns, and being out of the normal state. It means finding the data of.

As shown in FIG. 1, the abnormality detection device 1 of the present embodiment detects an abnormality from changes in signals from various sensors, for example, and is a signal input unit 2, an A / D conversion unit 3, and a data storage unit. 4. The identification unit 5 is provided and is roughly configured.

The signal input unit 2 inputs analog signals from various sensors such as a microphone, a vibration sensor, and an optical sensor as measured signals and outputs them to the A / D conversion unit 3.

The signal to be measured in the present embodiment is not only a regular signal such as a sine wave or a square wave, but also a signal such as sudden mechanical vibration, a lightning surge, or noise, irregular noise, distortion, or the like. Includes the signal of.

The A / D conversion unit 3 converts the signal to be measured from the signal input unit 2 into digital data and outputs it to the data storage unit 4.

The data storage unit 4 stores the digital data of the signal to be measured that has been A / D converted by the A / D conversion unit 3.

The identification unit 5 discriminates whether the digital data of the signal to be measured stored in the data storage unit 4 is normal or abnormal, and is a digital signal processor 11, a provisional determination device 12, a machine learning classifier 13, and a classifier. It includes a tuning data update unit 14, a classifier tuning data storage unit 15, a teacher data update unit 16, a teacher data storage unit 17, and a machine learning classifier update unit 18.

The digital signal processor 11 performs a process of extracting a plurality of types of pre-designed feature quantities from the digital data of the signal to be measured stored in the data storage unit 4.

Each feature amount extracted by the digital signal processor 11 is a numerical value of what kind of features the digital data of the signal to be measured has, as shown in (1) to (4) below, for example. , Various things are designed in advance by hand.

(1) Only the signal component showing a steep change is extracted from the digital data of the signal to be measured by the digital filter (bandpass filter), and the peak amplitude value of the extracted signal component is extracted as the feature amount of the signal to be measured.
(2) A singular point of the frequency of the digital data of the signal to be measured is extracted as a feature amount of the signal to be measured by a fast Fourier transform (FFT).
(3) The change in the slope of the digital data of the signal to be measured is extracted by the differentiator as the feature amount of the signal to be measured.
(4) The median filter detects the central amplitude value detected by comparing three adjacent points from the digital data of the signal to be measured, and the amplitude value obtained by subtracting the detected central amplitude value from the original amplitude value is applied. It is extracted as a feature amount of the measurement signal.

In the provisional determination device 12, identification boundary values (threshold values) are set in advance for each of a plurality of types of feature amounts, and identification boundary values (identification boundary values) corresponding to each feature amount of the signal to be measured extracted by the digital signal processor 11 ( A threshold value) is compared, and based on the comparison result, a process of tentatively determining whether the digital data of the signal to be measured is normal or tentatively abnormal is performed for each feature amount. The "temporary abnormality" is a state in which it is not normal, and means a state in which it is doubtful whether the digital data of the signal to be measured is normal or abnormal.

The machine learning classifier 13 has a function of setting an identification boundary value (threshold value) according to the digital data of the signal to be measured stored in the data storage unit 4, and is a methodology for giving a computer a learning ability. By machine learning, it is determined whether the digital data of the signal to be measured, which is provisionally determined to be a provisional abnormality (not normal) by the provisional determination device 12, is normal or abnormal.

The identification boundary value (threshold value) of the initial state of the machine learning classifier 13 is set using the digital data of the signal to be measured. Specifically, the teacher data of the teacher data storage unit 17 updated by the teacher data update unit 16 based on the normally labeled classifier tuning data previously stored in the classifier tuning data storage unit 15. The identification boundary value (threshold) of the initial state is set by the machine learning classifier update unit 18 using.

As the machine learning classifier 13, for example, a support vector machine (SVM), a neural network, a convolutional neural network, or the like is used.

The support vector machine is an algorithm that finds the optimum identification boundary value (threshold value) for separating two classes of digital data (normal data and abnormal data).

A neural network is a model that imitates the mechanism of the neural circuit of the human brain, and various problems can be solved by learning with a computer. However, know-how is required for model design. In addition, although it is difficult to take time to learn when multiple layers are used, a multi-layered model has appeared due to improvements in computing power, evolution of learning algorithms, and effective use of large amounts of digital data.

A convolutional neural network is composed of layers having characteristic functions such as a convolutional layer, and is particularly effective in image recognition. The convolution layer has a role of feature extraction such as edge extraction in an image.

The classifier tuning data update unit 14 is a machine learning classifier 13 likelihood (output value) of the classifier tuning data with a normal label of the classifier tuning data storage unit 15 and a provisional determination unit 12. Based on the comparison result of the digital data of the measured signal tentatively determined to be normal with the likelihood (output value) of the machine learning classifier 13, a normal label stored in the classifier tuning data storage unit 15 is attached. Update the classifier tuning data.

Further explaining, the classifier tuning data update unit 14 is among the likelihoods (output values) of the machine learning classifier 13 of the classifier tuning data with the normal label of the classifier tuning data storage unit 15. The maximum value is compared with the likelihood (output value) of the machine learning classifier 13 of the digital data of the signal to be measured that is tentatively determined to be normal by the tentative determination device 12, and the measurement is tentatively determined to be normal by the tentative determination device 12. If the likelihood (output value) of the digital data of the signal is larger, the digital data of the signal to be measured that is tentatively determined to be normal by the tentative judge 12 is used as the classifier tuning data with a normal label for classifier tuning. The data is additionally saved in the data storage unit 15 and the data is updated.

The classifier tuning data storage unit 15 labels the digital data of the signal to be measured as normal or abnormal based on the judgment result of the provisional judgment device 12 or the judgment result by machine learning of the machine learning classifier 13. Data is stored in memory. The classifier tuning data stored in the memory of the classifier tuning data storage unit 15 is updated by the classifier tuning data update unit 14.

The teacher data update unit 16 sets the identification boundary value in which the likelihood (output value) of the machine learning classifier 13 is preset among the classifier tuning data with the normal label of the classifier tuning data storage unit 15. The classifier tuning data that is equal to or greater than the (threshold value) is additionally stored in the teacher data storage unit 17 as teacher data with a normal label, and the data is updated.

In addition, the teacher data update unit 16 classifier tuning with a predetermined number (for example, about 1/10 of the distributed data) in descending order of likelihood (output value) of the machine learning classifier 13. The data for use is additionally stored in the teacher data storage unit 17 as teacher data with a normal label, and the data is updated.

The teacher data storage unit 17 has a memory for updating the machine learning classifier 13, and digital data of the signal to be measured based on the determination result of the provisional determination device 12 or the determination result by machine learning of the machine learning classifier 13. Store teacher data labeled as normal or abnormal. The teacher data stored in the memory of the teacher data storage unit 17 is updated by the teacher data update unit 16.

The machine learning classifier update unit 18 updates the identification boundary value (threshold value) of the machine learning classifier 13 using the teacher data with the normal label updated by the teacher data update unit 17.

Next, an abnormality detection method using the abnormality detection device configured as described above will be described. In order to detect an abnormality in the signal to be measured, the first learning course described later is executed. Then, when the measured signals from various sensors are input to the signal input unit 2, the measured signals are converted into digital data by the A / D conversion unit 3 and then stored in the data storage unit 4.

Next, the digital signal processor 11 extracts a plurality of types of pre-designed feature quantities from the digital data of the signal to be measured stored in the data storage unit 4.

Next, the provisional determination device 12 provisionally determines whether or not the signal to be measured is normal by comparing each feature amount extracted by the digital signal processor 11 with the identification boundary value (threshold value) corresponding to each individual. .. The tentative determination device 12 outputs a normal determination signal when the measured signal is tentatively determined to be normal, and outputs a tentative abnormality determination signal when the measured signal is tentatively determined to be tentative.

Next, the machine learning classifier 13 sets the identification boundary value (threshold) according to the digital data of the signal to be measured, and when the provisional abnormality determination signal is input from the provisional determination device 13, it is provisionally determined as a provisional abnormality. The digital data of the signal to be measured is machine-learned to identify whether it is normal or abnormal. The machine learning classifier 13 determines that the signal is abnormal if the likelihood (output value) of the digital data of the signal to be measured that is temporarily determined to be tentatively abnormal is equal to or higher than the specified value, that is, equal to or higher than the latest discrimination boundary value (threshold value). An abnormality judgment signal is output, and if it is less than the latest identification boundary value (threshold value), it is judged to be normal and a normal judgment signal is output.

Further, the operation of machine learning by the machine learning classifier 13 will be described by dividing it into a first learning course, a second learning course, a third learning course, and an operation process with reference to FIGS. 2 to 5. In performing machine learning, as a preliminary preparation, the digital data of the signal to be measured including the normal and abnormal parts is acquired, the digital data of the signal to be measured at the normal part is labeled as normal, and the abnormal part is attached. Label the digital data of the signal under test at the location as abnormal. Then, a part of the digital data labeled as normal and abnormal is taken out and stored in the data storage unit 4 as learning data, and the rest is stored in the classifier tuning data storage unit 15 as tuning data.

[First study course]
The first learning course shown in FIG. 2 is a course for setting the discrimination boundary value (threshold value) of the machine learning classifier 13. In the first learning course, a learning condition, a convergence condition, a number of repetitions, a learning rate, and the like are set as learning stop conditions (ST1). Then, the learning data stored in the data storage unit 4 is used as the first learning data, and machine learning is performed using the first learning data (ST2). Then, it is determined whether or not the learning stop condition is satisfied (ST3), and when it is determined that the learning stop condition is satisfied (ST3-Yes), the first learning course is terminated. If it is determined that the learning stop condition is not satisfied (ST3-No), machine learning is performed again using the learning data stored in the data storage unit 4 after changing the condition (ST2). As a result of this machine learning, the discrimination boundary value (threshold value) is set using the teacher data with a normal level.

[Second study course]
The second learning course shown in FIG. 3 is a course for updating the machine learning classifier 13. In the second learning course, the digital data of the signal to be measured that is determined to be normal is regarded as normal data, and machine learning is performed using this normal data (ST11). Then, it is determined whether or not the likelihood (output value) of the machine-learned normal data is equal to or higher than the latest identification boundary value (threshold value) (ST12). When it is determined that the likelihood (output value) is equal to or higher than the identification boundary value (threshold value) (ST12-Yes), the classifier tuning data update unit 14 stores the normal data as the classifier tuning data as the classifier tuning data. It is additionally stored in the database of the part 15 (ST13). When it is determined that the likelihood (output value) is not equal to or higher than the discrimination boundary value (threshold value) (ST12-No), the second learning course is terminated.

[Third study course]
The third learning course shown in FIG. 4 is a course for updating the discrimination boundary value (threshold value) of the machine learning classifier 13. In the third learning course, a learning condition, a convergence condition, a number of repetitions, a learning rate, and the like are set as learning stop conditions (ST21). Further, the teacher data update unit 16 updates the database of the teacher data storage unit 17 based on the tuning data of the classifier tuning data storage unit 15 additionally stored in the second learning course, and the updated teacher data is stored. With the learning data of 2, the machine learning classifier updating unit 18 updates the identification boundary value (threshold) of the machine learning classifier 13 based on the second learning data (ST22). Then, the updated second learning data is machine-learned (ST23), it is determined whether or not the learning stop condition is satisfied (ST24), and it is determined that the learning stop condition is satisfied (ST24-Yes). Complete the third study course. When it is determined that the learning stop condition is not satisfied (ST24-No), the teacher data (second learning data) stored in the teacher data storage unit 17 is machine-learned again by changing the condition (ST23).

[Operation process]
The operation process shown in FIG. 5 is a process of discriminating whether the digital data of the signal to be measured, which is provisionally determined to be a provisional abnormality by the provisional determination device 12, is normal or abnormal by machine learning. In the operation process, the digital data of the signal to be measured that is provisionally determined to be a temporary abnormality is used as inspection data, and the likelihood (output value) output from the machine learning classifier 13 by inputting this inspection data is the latest identification boundary value ( It is determined whether or not it is equal to or greater than the threshold value (ST31). Then, if the likelihood (output value) is equal to or greater than the identification boundary value (threshold value) (ST31-Yes), it is determined to be “abnormal” (ST32). If the likelihood (output value) is not equal to or greater than the identification boundary value (threshold value) (ST31-No), it is determined to be "normal" (ST33).

By the way, in the present embodiment, as shown in FIG. 1, the signal to be measured (analog signal) input from various sensors to the signal input unit 2 is converted into digital data by the A / D conversion unit 3. However, when the digital data of the signal to be measured is input to the signal input unit 2, the configuration of the A / D conversion unit 3 can be omitted.

Further, the components (signal input unit 2, A / D conversion unit 3, data storage unit 4, and identification unit 5) included in the abnormality detection device 1 are composed of a computer equipped with an arithmetic processing unit, a storage device, and the like. , The processing of each component may be executed by the program.

As described above, in the present embodiment, for the signals to be measured from various sensors, a plurality of types of feature quantities designed in advance are extracted, and each extracted feature quantity is compared with the corresponding identification boundary value (threshold value). Based on this, the signal to be measured is tentatively determined as normal or tentatively abnormal. Then, the measured signal that is tentatively determined to be tentatively abnormal is classified into normal and abnormal by data analysis by statistical modeling using machine learning. As a result, even if the signal to be measured from various sensors is a regular signal, a sudden signal, or an irregular signal, it is possible to detect an abnormality more accurately than before without depending on the analysis target.

Further, when the present embodiment is applied to, for example, an article inspection on a production line, it is possible to update the identification boundary value (threshold value) of the machine learning classifier while operating the machine learning classifier to detect the presence or absence of an abnormality in the signal to be measured. it can. Moreover, after the signal to be measured that is suspected to be normal or abnormal is tentatively determined as a tentative abnormality, the final determination of the signal to be measured that is tentatively determined to be a tentative abnormality is performed using machine learning. The rate of erroneous judgment of normal as abnormal is reduced in the part. As a result, even if the identification boundary value (threshold value) for each feature amount is loosely set to improve the abnormality detection accuracy as in the conventional case, the rate of erroneously determining a non-defective product that should be judged as normal is reduced, and a non-defective product is reduced. There is less waste, and there is no risk that the yield of products will deteriorate and production efficiency will drop significantly.

[Application example]
The anomaly detection device, the anomaly detection method, and the anomaly detection program of the present embodiment are within the normal operation range by statistical modeling technology, for example, when detecting the vibration of a manufacturing line or processing equipment and predicting an abnormality (failure). When detecting machine abnormalities by determining whether or not, when recognizing machine operation sounds, daily life sounds such as opening and closing of doors, and environmental sounds, when detecting belt conveyor vibration and detecting bearing wear, windows and It can be applied when detecting suspicious movements such as picking by detecting the seismic intensity of the door, when detecting the sound of a car and detecting mischief or theft, and when detecting anomalies in the signal to be measured. ..

The best form of the anomaly detection device, the anomaly detection method, and the anomaly detection program according to the present invention has been described above, but the present invention is not limited by the description and drawings in this form. That is, it goes without saying that all other forms, examples, operational techniques, and the like made by those skilled in the art based on this form are included in the scope of the present invention.

1 Anomaly detection device 2 Signal input unit 3 A / D conversion unit 4 Data storage unit 5 Identification unit 11 Digital signal processor 12 Temporary judgment unit 13 Machine learning classifier 14 Classifier Tuning data update unit 15 Classifier Tuning data storage Department 16 Teacher Data Update Department 17 Teacher Data Storage Department 18 Machine Learning Classifier Update Department

Claims (3)

A data storage unit (4) that stores digital data of the signal to be measured, and
An abnormality detection device (1) including an identification unit (5) for identifying whether the digital data is normal or abnormal.
The identification unit
A digital signal processor (11) that extracts a predetermined feature amount of the digital data stored in the data storage unit, and a digital signal processor (11).
A provisional determination device (12) that provisionally determines whether or not the digital data is normal based on the extraction result of the feature amount by the digital signal processor, and
A machine learning classifier (13) that discriminates whether digital data that is provisionally determined to be normal by the provisional determination device is normal or abnormal by machine learning , and
The teacher data storage unit (17), which stores teacher data labeled as normal or abnormal, and
A classifier tuning data storage unit (15) that stores classifier tuning data labeled as normal or abnormal, and a classifier tuning data storage unit (15).
The output value of the machine learning classifier of the classifier tuning data with the normal label of the classifier tuning data storage unit and the machine learning classifier of the digital data tentatively determined to be normal by the tentative judgment device. A classifier tuning data update unit (14) that updates the normally labeled classifier tuning data stored in the classifier tuning data storage unit based on the comparison result with the output value, and a classifier tuning data update unit (14).
Among the updated normal labeled classifier tuning data of the classifier tuning data storage unit, the classifier tuning data in which the output value of the machine learning classifier is equal to or higher than a preset threshold value is normal. A teacher data update unit (16) that is additionally stored in the teacher data storage unit as teacher data labeled with
An abnormality detection device including a machine learning classifier update unit (18) that updates the machine learning classifier using teacher data with a normal label updated by the teacher data update unit. It is an abnormality detection method that stores the digital data of the signal to be measured and identifies whether the digital data is normal or abnormal.
A step of extracting a predetermined feature amount from the stored digital data, and
A step of tentatively determining whether or not the digital data is normal based on the extraction result of the feature amount, and
A step of identifying whether the digital data tentatively determined to be normal by the tentative determination process is normal or abnormal by machine learning , and
Steps to store teacher data labeled normal or abnormal, and
Steps to store classifier tuning data labeled normal or abnormal, and
Normal label based on the comparison result between the output value of the classifier tuning data with a normal label by the machine learning and the output value of the digital data tentatively determined to be normal by the tentative determination process by the machine learning. Steps to update the classifier tuning data with
Of the updated classifier tuning data with a normal label, the classifier tuning data whose output value by machine learning is equal to or higher than a preset threshold value is additionally stored as teacher data with a normal label. Steps and
An anomaly detection method comprising the step of updating the machine learning using the updated teacher data labeled as normal. Computer,
A data storage unit (4) that stores digital data of the signal to be measured, and
A digital signal processor (11) that extracts a predetermined feature amount of the digital data stored in the data storage unit, and a digital signal processor (11).
A provisional determination device (12) that provisionally determines whether or not the digital data is normal based on the extraction result of the feature amount by the digital signal processor, and
A machine learning classifier (13) that discriminates whether digital data that is provisionally determined to be normal by the provisional determination device is normal or abnormal by machine learning , and
The teacher data storage unit (17), which stores teacher data labeled as normal or abnormal, and
A classifier tuning data storage unit (15) that stores classifier tuning data labeled as normal or abnormal, and a classifier tuning data storage unit (15).
The output value of the machine learning classifier of the classifier tuning data with the normal label of the classifier tuning data storage unit and the machine learning classifier of the digital data tentatively determined to be normal by the tentative judgment device. A classifier tuning data update unit (14) that updates the normally labeled classifier tuning data stored in the classifier tuning data storage unit based on the comparison result with the output value, and a classifier tuning data update unit (14).
Among the updated normal labeled classifier tuning data of the classifier tuning data storage unit, the classifier tuning data in which the output value of the machine learning classifier is equal to or higher than a preset threshold value is normal. A teacher data update unit (16) that is additionally stored in the teacher data storage unit as teacher data labeled with
An abnormality detection program for functioning as a machine learning classifier update unit (18) that updates the machine learning classifier using the teacher data with a normal label updated by the teacher data update unit.

Images